Brain Extraction Network Trained with "Silver Standard" Data and Fine-Tuned with Manual Annotation for Improved Segmentation
Roberto Medeiros de Souza, Oeslle Lucena, Mariana Bento, Julia Lopes Garrafa, Letícia Rittner, Simone Appenzeller, Roberto Lotufo, Richard Frayne · 2019
Training convolutional neural networks (CNNs) for medical image segmentation often requires large and representative sets of images and their corresponding annotations. Obtaining annotated images usually requires manual intervention, which is expensive and time consuming, as it typically requires a specialist. An alternative approach is to leverage existing automatic segmentation tools and combine them to create consensus-based "silver-standards" annotations. A drawback to this approach is that silver-standards are usually smooth and this smoothness is transmitted to the output segmentation of the network. Our proposal is to use a two-staged approach. First, silver-standard datasets are used to generate a large set of annotated images in order to train the brain extraction network from scratch. Second, fine-tuning is performed using much smaller amounts of manually annotated data so that the network can learn the finer details that are not preserved in the silver-standard data. As an example, our two-staged brain extraction approach has been shown to outperform seven state-of-the-art techniques across three different public datasets. Our results also suggest that CNNs can potentially capture inter-rater annotation variability between experts who annotate the same set of images following the same guidelines, and also adapt to different annotation guidelines.